Dual-Task Learning for Dead Tree Detection and Segmentation with Hybrid Self-Attention U-Nets in Aerial Imagery

TL;DR

This work tackles dead-tree detection in high-resolution aerial imagery, addressing challenges posed by dense canopies and spectral similarity between live and dead vegetation. It introduces TreeMort-3T-UNet, a multi-task architecture with segmentation, centroid localization, and a hybrid SDT-boundary map, trained with a hybrid loss and refined by a watershed-based postprocessing pipeline. On a boreal Finland dataset, it achieves substantial gains in instance-level segmentation (Mean Tree IoU up ~41%) and centroid localization (centroid error reduced ~57%), demonstrating robustness in dense canopy conditions. The approach supports scalable wall-to-wall mortality mapping with practical impacts on wildfire risk assessment, carbon stock estimation, and precision forestry.

Abstract

Mapping standing dead trees is critical for assessing forest health, monitoring biodiversity, and mitigating wildfire risks, for which aerial imagery has proven useful. However, dense canopy structures, spectral overlaps between living and dead vegetation, and over-segmentation errors limit the reliability of existing methods. This study introduces a hybrid postprocessing framework that refines deep learning-based tree segmentation by integrating watershed algorithms with adaptive filtering, enhancing boundary delineation, and reducing false positives in complex forest environments. Tested on high-resolution aerial imagery from boreal forests, the framework improved instance-level segmentation accuracy by 41.5% and reduced positional errors by 57%, demonstrating robust performance in densely vegetated regions. By balancing detection accuracy and over-segmentation artifacts, the method enabled the precise identification of individual dead trees, which is critical for ecological monitoring. The framework's computational efficiency supports scalable applications, such as wall-to-wall tree mortality mapping over large geographic regions using aerial or satellite imagery. These capabilities directly benefit wildfire risk assessment (identifying fuel accumulations), carbon stock estimation (tracking emissions from decaying biomass), and precision forestry (targeting salvage loggings). By bridging advanced remote sensing techniques with practical forest management needs, this work advances tools for large-scale ecological conservation and climate resilience planning.

Figures (4)

• Figure 1: The TreeMort-3T-UNet architecture for multi-task learning in instance segmentation. A pre-trained ResNet-34 encoder extracts robust features from multispectral (RGB-NIR) aerial imagery. The hierarchical decoder, connected via skip connections, bifurcates into three output branches: a segmentation head that produces a binary mask, a centroid head that outputs a Gaussian heatmap for instance localization, and a hybrid head that generates a signed distance transform (SDT) combined with explicit boundary cues. Self-Attention modules refine feature maps to handle overlapping and densely clustered tree canopies effectively.
• Figure 2: Visualizing the dataset and its spatial characteristics. From left to right: Spatially aware partitioning of the dataset into training (70%), validation (20%), and test (10%) sets, ensuring spatial independence and proportional distribution of dead tree segments across Finland. The partitions are color-coded and overlaid on a map of Finland, with geographic regions clustered based on latitude and longitude bins. Heatmap illustrating the density of dead tree centroids across the dataset, highlighting areas of concentrated tree mortality. A representative aerial image from the dataset, showing a sample plot. A black rectangle on the Finland map indicates the approximate location of this sample plot. A zoomed-in region of the sample plot, with dead tree annotations outlined in red polygons. Binned histogram of dead tree segment sizes, providing insights into the size distribution of mortality segments. The segment compactness distribution depicts the shape characteristics of the detected dead tree regions. Compactness is defined as $\frac{4\pi \times \texttt{area}}{\texttt{perimeter}^2}$, with higher values indicating shapes that are more circular.
• Figure 3: Sample images showcasing the segmentation process. (a) Input aerial image with ground truth annotation, (b) Model’s initial baseline prediction, and (c) Final postprocessed segmentation.
• Figure 4: Paired scatter plot comparing Mean Pixel IoU and Mean Tree IoU across different model configurations. Each data point represents a specific model's performance, highlighting the relationship and potential trade-offs between pixel-level accuracy and instance-level segmentation precision.